Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.



Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

Man I have been dying to make this one for a while now. First rumours for RDNA3 are here so new thread time! Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3...

forums.anandtech.com

 

[Jan Olšan]

AMD Radeon RX 7900 GRE review: the lowest binned Navi 31 variant is now globally available, starting at $549

Navi 31 hops on a golden rabbit.

www.tomshardware.com

It launched globally at $549

​

​

Yeah but $549 is likely too low.

The $649 might be it. I think the idea that they are comparing with 7900 GRE because the launch price will be the same (original launch price, that is how it is always done), makes a lot of sense. That doesn't guarantee that the prediction is perfect, but not totally bad odds, IMHO.

 

[linkgoron]

-But it was a long way off from the 2080Ti.

As was stated in my quote, but it was just 250mm^2. Just creating a large die would've given them a huge performance enhancement, without a new arch.

- The fact that RDNA2 was stuck on the same node and had relatively narrow bus widths did it for a lot of people. Infinity Cache really came out of nowhere.

Also, let us please not ignore the deep persistent irony of RDNA3, which ended up with 500mm^2+ of RDNA3 breaking even against 370mm^2 of Ada (and losing in RT)... exactly the argument that was supposedly outlandish when talking about RDNA2...

First of all - I'll start off by saying that RDNA3 is a turd. However, your comparison not exactly equivalent. AD103 is a monolithic 379mm^2 5nm GPU. I'm not sure I'd call RDNA3 a "real" 500mm^2 card. It's a mix of two different processes (6nm which is a 7nm derivative and 5nm). The GCD is ~304mm^2 and the memory chiplets are ~225=37.52*6, and I'm not sure how much the chiplet architecture and memory parts increase the die size, vs a monolithic design (for both the GCD and MCD). I can assume that it increases the die size. Also, GCD density is three times higher than MCD density (150M/mm^2 vs 50M/mm^2 although that might also be a characteristic of what they do and not just the process. For the record, AD103 is 121M/MM^2), so that's a clear sign IMO that monolithic NAVI 31 would have been much smaller.

- The whole argument was "it's silly to think a 500mm2 AMD part would compete with a 390mm^2 Nvidia part" and yet that's exactly what happened.

Retroactive precedent

You don't need RDNA3 for that as a precedent - we have RX Vega 64 with a GPU at 495 mm², barely competing with 314 mm² GTX 1080. However, looking at 5700xt RDNA1 being 250mm^2, a new card with a continuation of RDNA2 over RDNA1 (which was a successful arch) at 500mm² would have never been just 50% faster.

Even the 7900xtx, which we all agree is a turd, is about 50% better than the 6900xt (in 4k), and only the GCD is a shrink, and it's using chiplets.

 

[Jaskalas]

Not happening. AMD is between a rock and a hard place right now. 5070Ti turned out to be fairly potent, matching 4080 Super.

Potent?

In years past, we expected the 60 class card to match the previous 80 class. And they did. For $400 or less.
Now it costs us $900 for that class of performance. 50 series is seriously underwhelming. Should be easy for AMD to give us a better value.

 

[linkgoron]

Potent?

In years past, we expected the 60 class card to match the previous 80 class. And they did. For $400 or less.
Now it costs us $900 for that class of performance. 50 series is seriously underwhelming. Should be easy for AMD to give us a better value.

Amazing that anyone would call the next gen lower step equaling the previous gen "upper step" potent. We've really thrown all expectations out of the window. What was the other option? 5070ti meeting the 4070ti Super?

For the record 4070ti matched the 3090 at 4k, the 3070ti beat the 2080ti by 10%, and even the 2070 beat the 1080 by 20%.

 

[branch_suggestion]

Careful with comparing a percentage from one source to others. It can have up to double digits error margin because games are varied that much (and that's putting aside possible cherry picking and other distortions to the original percentage).

I tried to fit the leaked numbers to ComputerBase and TPU and ended up with 9070XT possibly being anywhere between 88-98% of RTX 5070 Ti performance. It looks like XT won't be faster/matched with it, that's probably certain now.

Eh, it depends, most games could be 5% faster on average, but there are a handful of membw bound games that it will lose to the XTX/5070Ti in at 4K.
I bet it will be faster at 4k raster in a full geomean of reviewers.

 

[Elfear]

Even the 7900xtx, which we all agree is a turd, is about 50% better than the 6900xt (in 4k), and only the GCD is a shrink, and it's using chiplets.

Wait, when did this happen?

 

[Tuna-Fish]

That's not how tariffs work. AMD makes the same amount of money and the consumer pays the additional cost of the tariff. It's no different than any other tax in that way. Similarly, AMD doesn't make more money holding off sales until a VAT increase went into effect for example.

Tariffs are levied at the time goods are imported, not at the time when they are sold. You do not have to pay tariffs for any goods imported before the tariffs come into effect, even if you sell them after that.

The general price level of electronics are going to go up early next month because tariffs come into effect for everything. If AMD currently has warehouses full of GPUs in the US, they can just hold on to them for a month and hope to sell them at a higher price. This would be a one-time windfall.

 

[Ghostsonplanets]

So N44 lacks Media Engine and has PCIe x4? Previously I was under the impression that a 9060 would be based on N44. But the lack of Media Engine and x4 would be a odd regression from N33 and N23 even. So I'm guessing AMD plan is to 9060 to be a N48 cutdown while they'll release a 32 CU 8GB 9050 and a 6GB 28CU 9040.

This possible also explains the previous rumor of N44 being ~130mm². Regardless, such line-up would be extremely competitive in the <$200 market.

 

[branch_suggestion]

Spoiler: NSFW LANGUAGE

https://twitter.com/x/status/1893264118573846627

AMD is asking reviewers for pricing suggestions.
They have never done this before pricing is set, only feedback afterwards.

 

[Ghostsonplanets]

https://twitter.com/x/status/1893174845703143506

Navi 48 is smaller than the 390mm² previous measurement. Smaller than GB203 (378mm²)

 

[Kepler_L2]

So N44 lacks Media Engine and has PCIe x4? Previously I was under the impression that a 9060 would be based on N44. But the lack of Media Engine and x4 would be a odd regression from N33 and N23 even. So I'm guessing AMD plan is to 9060 to be a N48 cutdown while they'll release a 32 CU 8GB 9050 and a 6GB 28CU 9040.

This possible also explains the previous rumor of N44 being ~130mm². Regardless, such line-up would be extremely competitive in the <$200 market.

It's probably x8 PCIe since N48 is x16. And yeah it seems to lack media encoder.

 

[Tuna-Fish]

In the chip segment slide, there is a sizable gap between parts of the market covered by 48 and 44, while the RDNA3 side was a solid block.

Possibly it's only designed to match 7600 at a lower power consumption level, not beat it. 4x PCIe 5.0 would be more than sufficient for the performance, but it would be kind of hard to justify 60 series naming.

 

[Kepler_L2]

In the chip segment slide, there is a sizable gap between parts of the market covered by 48 and 44, while the RDNA3 side was a solid block.

Possibly it's only designed to match 7600 at a lower power consumption level, not beat it. 4x PCIe 5.0 would be more than sufficient for the performance, but it would be kind of hard to justify 60 series naming.

RDNA4 has nearly 50% improved performance per CU.

 

[IEC]

https://twitter.com/x/status/1893264118573846627

AMD is asking reviewers for pricing suggestions.
They have never done this before pricing is set, only feedback afterwards.

That's encouraging that AMD is asking for feedback on pricing. Hopefully they listen to Steve from HWUB - aggressive undercutting with high supply would win them a LOT of mindshare right now. Especially because anything over RTX 4060 level is basically unobtanium or stupidly priced.

 

[jpiniero]

That's encouraging that AMD is asking for feedback on pricing. Hopefully they listen to Steve from HWUB - aggressive undercutting with high supply would win them a LOT of mindshare right now. Especially because anything over RTX 4060 level is basically unobtanium or stupidly priced.

I don't think they can. Well, if AMD wants to actually make money.

 

[gaav87]

I don't think they can. Well, if AMD wants to actually make money.

Right... cause selling 1milion units for 699$ would be better than 3-4mil at 549 or 6-7mil at 499$

 

[gaav87]

High supply would make AIB happy and maybe they would get some OEM in it not only diy.

 

[jpiniero]

Right... cause selling 1milion units for 699$ would be better than 3-4mil at 549 or 6-7mil at 499$

"We'll lose money... but we'll make it back up in volume"

 

[basix]

"We'll lose money... but we'll make it back up in volume"

As the 7800XT was priced at 499$ you could also price a 9070 XT at 499$ and still make money. But I think 599$ for 9070XT and 499$ for the 9070 are more likely to happen. Maybe even a little bit higher.

A 9070 at 499$, with 16GB and with raster performance on ~4070 TiS level would provide very decent bang for the buck. And even more so, if in stores and buyable for that amount of money.
Competetivness with RT and FSR4 vs. DLSS4 ist to be seen but even if RDNA4 lags behind a bit it would still be a good deal compared to whatever is on the market right now.

Especially 9070 non-XT looks like it was designed to be a "bang for the buck king" (salvage, low TDP, small cards, maybe lower speced GDDR6 at 18Gbps) and that is exactly what gamers long for at the moment

 

[gdansk]

At $699 or higher they'll only sell while Nvidia has supply issues. Possibly through April if MCK was correct.
7000 series showed how little cutting prices after launch changed perception.
The 7900 XT is more expensive to make and needs more VRAM chips. It sold for $625-$640 frequently. With its innovative disaggregated design it surely had even more development costs to recover.
Radeon division has been profitable, even if only slightly, the entire time it has been on the market even with only 10% market share.

I'm pretty sure AMD has the option to reduce the MSRP for perception if they want to. Especially if they copy Nvidia's fake MSRP strategy. But there are also scenarios it makes sense not to price aggressively. E.g. somehow it is faster than chopped GB203 (it isn't) or if AMD simply doesn't want to waste N4 allocation on a reasonably large chip (probable).

 

[Josh128]

https://twitter.com/x/status/1893264118573846627

AMD is asking reviewers for pricing suggestions.
They have never done this before pricing is set, only feedback afterwards.

If AMD emailed HUB for pricing suggestions and HUB responded with $549 max for XT, AMD likely deleted that email, lol. They are not going to like the feedback they get on such a question. Besides, that question is impossible to answer with any degree of seriousness without knowing the approximate manufacturing costs.

 

[Josh128]

RDNA4 has nearly 50% improved performance per CU.

Vs RDNA 3? Thats some incredible hopium. They probably wont even get that in ray tracing, let alone raster.

 

[maddie]

Vs RDNA 3? Thats some incredible hopium. They probably wont even get that in ray tracing, let alone raster.

That's what it takes to get between XT and XTX models of 7900.

 

[gaav87]

Navi48 confirmed at 350mm2

 

[PJVol]

Thats some incredible hopium

96 rdna3 cu's / 64 rdna4 cu's - clocks improvement = "nearly 50%".
Just simple, Annalena B. 360° math.